🗣️ 19. The lexicon: layers, borrowings, multiword expressions and register

Brunot describes French vocabulary as three layers (§165): a popular fund inherited from spoken Latin and worn by fifteen centuries of sound change (eau < aqua, père < patrem), a learned fund taken from written Latin and Greek and kept close to its source (aquatique, paternel), and a fund of foreign origin. English took much of its learned and Norman vocabulary from French, so the two languages share tens of thousands of words and many traps.

This chapter covers the layers and doublets (§§1–2), borrowings (§3), multiword expressions (§4), collocations (§5), the register of words (§6) and false friends with English (§7). §8 says what Mova has and lacks. Word formation is in chapters 16–18, the grammar of registers in 37, and the national varieties in 08. The Latin side of the layers is in the Latin manual, la 11.

We hold no French frequency dictionary and no etymological dictionary with counts, so this chapter gives no figures for the share of Latin, Germanic or English words in French. Its figures are counts on the UD 2.18 treebanks (2026-11-06) and results reported by the open papers cited.

1. The layers

The inherited fund holds the everyday words: body parts (main, œil, dent, pied), kinship (père, mère, frère, fils), the basic verbs (être, avoir, aller, faire, dire, voir, prendre), nature (pluie, mer, terre) and the grammatical words (le, de, et, que). Sound change has worn them: aqua became /o/, written eau; oculum became œil; the neuter plural folia became the feminine singular feuille. The inherited words are short, often irregular and highly polysemous.

Gaulish left very little. Brunot warns against the "celtomanes" and notes that many supposed Celtic sources are French words borrowed into Breton (just, coler). Since Gaulish shared many roots with Germanic, a word such as cloche may come from either. The words that Latin authors themselves cite as Gaulish are certain: alouette, cervoise, arpent, lieue. Others are attributed with less certainty (balai, bec, bruyère, cabane, grève, harnais), and they mostly came through Latin (brunot-1899, §142).

The learned fund came in from the twelfth century on, in waves, as law, theology, science and administration took from written Latin what the vernacular lacked (fragile, nocturne, digital), and from Greek, often through Latin (philosophie, théâtre, anatomie). It keeps Latin stems, which is why a French noun often has an adjective of another shape (§2.2). It is also the base of neoclassical compounding (bio-, géo-, -logie, -graphie, -phile; 17).

The foreign fund is divided by period (Germanic in the early Middle Ages, Italian in the Renaissance, English from the eighteenth century) and by register (recent Arabic and English words are familiar or slang). See §3.

What the layer of a word predicts:

2. Doublets and learned adjectives

2.1 Doublets

One Latin word often gave French two: an inherited one and a learned one (or one borrowed through another language), which the language then kept apart in meaning. Brunot's examples (§165) are basilicam, which gave popular basoche and learned basilique, and cantatam, which gave chantée directly and cantate through Italian cantata. He explains why the second word could enter at all. Sound change makes a word so different from its source (hôtel from hospitale, cheville from claviculam) that, once the etymology is forgotten, the Latin word can be borrowed again without anyone noticing that the language already has it. Likewise factionem gave French façon and English fashion, far enough apart to be borrowed back. Budget is Old French bougette "small purse" returning from England.

The last two show a second motive (§165). Naïf first meant "natural", as Latin nativus did, then "simple, gullible". When the pejorative sense took over, the old sense had to be taken again from Latin, as natif. Italian cadenza kept the literal "fall", which French chance ("what falls to one", luck) had lost. Once the mechanism exists, Brunot adds, it spreads to doublets nobody needed, such as a scholar's Latin form or a fashionable person's English word.

2.2 Learned adjectives to inherited nouns

The learned layer gives many inherited nouns an adjective of a different shape. The link is regular but cannot be derived from the noun, so a tool that looks for a suffix on the noun will not find it.

English works the same way (water → aquatic), and the learned word is usually the same in both languages. Derivational databases record such links as families rather than rules. Démonette2 (Namer et al. 2019, namer-2019) describes relations between lexemes, including indirect ones, in a network planned at no fewer than 366,000 entries. Sanacore, Hathout & Namer (2020, sanacore-2020) represent the meaning of a whole derivational family with morphosemantic frames.

3. Borrowings

3.1 The older layers

• Germanic. Old French had many Germanic words of every kind (war, seafaring, cooking; nouns, adjectives and verbs). It lost many of them, and their origin by period is often hard to fix: some date from the invasions, some came with the Normans, and some arrive "every day from England" (brunot-1899, §145). Examples are guerre, jardin, blanc, bleu, gris, riche, hardi, haïr, garder, gagner, honte, hache, trop. Frankish initial h is the source of the h aspiré that blocks liaison and elision (le haut, la honte; 04). Brunot shows how loans are made French: Frankish faldestôl gave fauteuil, marahskalk gave maréchal, Dutch bakboord gave bâbord, German Sauerkraut gave choucroute, English riding-coat gave redingote (§141). Later Dutch loans are mostly maritime (matelot, yacht, digue), and German ones include valse, chope, nouille, képi.
• Italian came above all in the sixteenth and seventeenth centuries, with the Italian wars and the court: balcon, banque, soldat, sonnet, caprice, concert, cantate, carnaval, façade, escalier. Brunot warns that borrowing did not stop there and lists modern ones, dilettante, imprésario, libretto, malaria, villégiature (§146); later still came spaghetti, pizza.
• Spanish, particularly in the seventeenth century: épagneul, guérilla, hâbler, jasmin, mantille, vanille, and camarade, sieste. Portuguese gave mandarin, caste, fétiche, autodafé (§147).
• Arabic, through Spanish, Italian and medieval Latin (science, trade, food): algèbre, chiffre, zéro, sucre, café, magasin, amiral, coton, matelas, carafe, jupe, hasard, sirop, tarif (§151). The colonial and post-colonial period brought direct and familiar loans: bled, toubib, kif, kiffer, maboul, flouze, chouïa, wesh.
• Other languages: kiosque, divan, bazar (Turkish, Persian); robot (Czech); judo, karaoké, manga, tsunami (Japanese); tomate, chocolat, cacao (Nahuatl through Spanish), maïs (Taíno).

3.2 Borrowings in a corpus: French rap

Zurbuchen & Voigt (2024, zurbuchen-2024) collected over 8,000 French rap songs through Spotify and Genius, with their years from 1991 to 2024, and annotated a lexicon of 741 borrowings by origin, part of speech and semantic class. Most of the lexicon comes from the English Wiktionary's category of French terms derived from other languages; the rest comes from words outside a French word list in 20 sampled songs and from embedding neighbours. In regressions on the yearly counts, shorter borrowings are clearly more used. Origin, semantic class and part of speech also play a part, and terms of music and the arts are the only semantic class with a positive effect in every model. Over time the share of all borrowings among the words of the lyrics roughly doubled since the 1990s, to about 2%. Niger-Congo loans rose fastest, Afroasiatic ones (many Arabic) stayed at about a tenth of the borrowings, and interjections rose after 2010 while nouns fell. The authors found automatic borrowing detection for French not robust enough and annotated by hand, and they note that the primary origin of a word was sometimes hard to fix. The corpus is a hotspot of borrowing, not a sample of French.

3.3 English

English loans begin in the eighteenth century (club, budget, partenaire, sport < Old French desport) and became a flood after 1945.

The 1990 rectifications give loans a French plural: des matchs, des solos, des lieds, des maximums. The singular is the most frequent form, even when that is a plural in the source language (un graffiti, des graffitis; un ravioli, des raviolis). Words still used as quotations are excluded (des mea culpa) (cslf-1990, rule II.7). Established foreign compounds are welded (weekend, cowboy, hotdog; 18 §3).

GSD marks foreign material with Foreign=Yes on 1,946 of its 400,385 words (0.49%, 2026-11-06). The commonest such forms are the (91), of (84), and (28) and de (24), all in titles and names of English works and companies. Integrated loans (week-end, parking, sandwich) are ordinary nouns with no Foreign feature, so it is the annotator who separates a quotation from a loan, not the form.

3.4 Language policy

These facts are public law, reported from general knowledge, not from a source in our library. The loi Toubon of 4 August 1994 requires French in advertising, consumer information, the workplace, public contracts and teaching. A 1996 decree set up terminology commissions, which propose French terms for fields where an English word has spread. Their terms are published in the Journal officiel and the FranceTerme database, and public bodies must use them. Examples are logiciel "software", ordinateur "computer" (proposed by a Latin professor, Jacques Perret, in 1955), baladeur "walkman", courriel "e-mail", mot-dièse "hashtag" and infox "fake news". The older ones succeeded; mot-dièse competes with hashtag. Québec's policy, through the Office québécois de la langue française, is older and more effective: courriel, clavardage, baladodiffusion, fin de semaine, stationnement, magasiner. Rap and youth speech are a main channel of new loans, and the Académie française resists many of them (Zurbuchen & Voigt 2024, introduction).

4. Multiword expressions

A multiword expression (MWE; expression polylexicale, locution) is a sequence of words with some idiosyncrasy of form, syntax, meaning or use, which a speaker stores and a program must treat as a unit. Pasquer (2017, pasquer-2017) recalls two often-quoted estimates: 40% of the words of Le Monde belong to MWEs (Gross & Senellart 1998), and a language has about as many MWEs as single words (Jackendoff 1997).

4.1 Types

The classes overlap: one sequence can be a collocation, a light-verb construction or an idiom depending on the analysis. The French PARSEME annotation uses a decision tree for the hard cases (Candito et al. 2017, candito-2017).

4.2 Annotation and variability

PARSEME French. Candito et al. (2017) annotated verbal MWEs in three main classes: idioms, inherently pronominal verbs and light-verb constructions (constructions à verbe support, a broadened notion). The corpus combines the French UD and Sequoia treebanks, 22,645 sentences with 4,962 expressions, about one per 100 tokens. Roughly 4 in 10 are idioms, 3 in 10 pronominal verbs and 1 in 3 light-verb constructions. 60.0% are continuous overall, but the figure is 76.6% for idioms, 85.8% for pronominal verbs and only 18.1% for light-verb constructions, so a light-verb construction is usually split by an article or an adjective. The same string can be literal or idiomatic (le roi ouvre alors les yeux sur son geste against la victime parle et ouvre les yeux). The MWE layer was released under CC BY 4.0, with the underlying morphosyntax under the licences of its source treebanks. The authors also record the choices of the earlier French Treebank: it marked faire face as an MWE but not avoir peur, because avoir une peur immense is possible.

Variation. Pasquer (2017) studied the 4,441 verbal MWE occurrences of the PARSEME French training corpus (17,880 sentences, 450,221 tokens; 2.2% of the tokens belong to a verbal MWE). 2,772 (62%) are continuous and 1,097 (25%) have a single inserted element, so no insertion or one insertion covers 87%. The single insertions differ by class. In pronominal verbs 79% are the auxiliary être (l'Empire s'est écroulé), in light-verb constructions 66% are articles, and in idioms 44% are adverbs and 28% articles. The noun of a light-verb construction varies in 82% of occurrences against 9.6% in idioms. The commonest idioms are avoir lieu (95) and faire partie (88). A lexicon of canonical forms therefore finds most occurrences but not all, and the paper's aim is to generate the variants.

Wordplay on expressions, défigement "unfreezing", changes a word while the expression stays recognisable, as in tu quoque mi fili → tu quoque mi chili. Bezançon et al. (2025, bezancon-2025) built FRUIT, 60,617 tweets from three years that contain MWEs and pun candidates. Three experts annotated 600 of them, with agreement α up to 0.83. A game with a purpose then gathered 2,206 annotations from 13 players on 931 tweets (α = 0.70), for 1,531 annotated tweets in all. They follow Mejri: the source expression must stay identifiable, or the pun fails.

4.3 Identification and understanding by machine

Bagdasarov, Alves & Teich (2026, bagdasarov-2026) compare a fine-tuned CamemBERT with two open generative models on the French data of PARSEME 2.0, which annotates MWEs of all categories, not only verbal ones. CamemBERT reaches F1 0.74 at the level of whole expressions, far above both LLMs. One of the LLMs gave unusable output (endless repetition or irrelevant text) in 128 of 373 cases. The LLMs' partial matches tend to contain too many tokens (faire confiance à for faire confiance) and CamemBERT's too few.

On understanding, EIFFEL (Noel et al. 2026, noel-2026) tests French idioms in context with input from native speakers. It shows the benefit of a higher share of French training data, which standard benchmarks do not show. Beauchemin et al. (2026, beauchemin-2026) build idiom benchmarks for Québec French (4,633 expressions and 171 regional words) and for Metropolitan French (4,938 expressions). Of the 111 models they test, 65.77% do significantly better on Metropolitan idioms and only 9.01% better on Québec ones.

For learners, PolylexFLE (Todirascu, Cargill & François 2019, todirascu-2019) has 4,295 verbal expressions, 1,098 of them labelled with a CEFR level by a mixed manual and automatic procedure.

4.4 What UD French annotates

UD marks only the grammatical end of the MWE scale: fixed for frozen function-word sequences, flat for names, compound, and the feature ExtPos for the part of speech of the whole. Idioms, light-verb constructions and spaced compounds get no UD mark and must come from PARSEME.

fixed is counted as dependents, so a three-word expression counts two. The spoken treebanks have about twice the density of the written ones. The commonest fixed expressions (head plus its fixed dependents, 2026-11-06) show the difference in register:

• GSD: ainsi que 128, en effet 102, à partir (de) 100, alors que 80 (+ alors qu' 43), un peu 75, par exemple 70, en tant que 59, d'ailleurs 58, tandis que 56, de plus 51, d'abord 49, en raison (de) 45, en particulier 43.
• ParisStories: en fait 153, parce que 140 (+ parce qu' 40), un peu 60, quand même 45, et tout 39, et puis 39, pas mal 15, une fois 12, en mode 6.

The connectives of writing (ainsi que, en effet, par exemple) give way in speech to discourse markers and hedges (en fait, quand même, et tout, en mode; 38). In GSD, ExtPos is ADV 1,942, PROPN 636, NOUN 482, PRON 429, SCONJ 318, CCONJ 311, ADP 308, DET 79, ADJ 55, VERB 18, INTJ 3 and NUM 2. In UD the first word of a fixed expression is its head, and ExtPos gives the class of the whole when it differs from the head's own.

5. Collocations

A collocation is a habitual combination in which one word, the base, is chosen for its meaning and the other, the collocate, is chosen by the base. Collocations lie between free combinations and idioms, and they differ between languages:

The verb for the same noun often differs from English (prendre une décision / make a decision; poser une question / ask a question), so such pairs are learned item by item. Meaning-Text theory names the relation between collocate and base as a lexical function: Magn for intensification (grièvement for blessé), Oper for the support verb (prendre for décision), Real for realisation. Collocations are found in corpora with association measures over dependency pairs (verb + obj, noun + amod). The light-verb constructions of §4 are the verb + object case.

6. Register

French dictionaries label words soutenu, littéraire, familier, populaire, argot, vulgaire, vieilli, régional, technique, enfantin, and a speaker chooses among near-synonyms by situation. The grammar of registers (questions, the dropping of ne, on for nous) is in 37. This section is about the choice of word.

6.1 Synonym series

The words move together with grammar: tu or vous, the question form and the omission of ne shift with them. The familiar register is marked by clipping (prof, resto) and -o (apéro), verlan (meuf, ouf, relou) and the other processes of 18; by the suffixes -ard, -asse, -oche (17); by metaphor (bagnole, caisse, citron "head"); by Arabic and English loans (§3); by intensifiers (vachement, super, grave; 15); and by fillers (bah, ben, genre, du coup, quoi, hein). The formal end has naguère, nonobstant, quérir, céans, the passé simple and imperfect subjunctive (13), inversion after peut-être and aussi, and ne alone (je ne saurais). Administrative French has its own fixed phrases (ci-joint, ci-dessus, aux fins de, ledit).

In the treebanks (form counts, 2026-11-06). Spoken ParisStories (42,789 words) has ça 309, bah 203, genre 82, truc 31, sympa 18 and cela 2. Written GSD, nearly ten times larger, has cela 142, ça 36 and truc 1. An example of the spoken style is genre, enfin, c'est ouf, genre, on dirait une peluche. (ParisStories_2021_loulouLeChat_1). A reader trained on GSD alone meets a different vocabulary in speech.

6.2 National and regional words

• Québec: magasiner, char "car", dépanneur "corner shop", bleuet, fin de semaine, chum "boyfriend", blonde "girlfriend".
• Belgium: septante, nonante, GSM "mobile phone", drache "downpour", s'il vous plaît for "here you are", dîner for the midday meal.
• Switzerland: septante, huitante (in some cantons), nonante, natel "mobile phone", poutzer "to clean", and the meals déjeuner "breakfast", dîner "lunch", souper "dinner".
• West Africa: s'enjailler "to have fun", une go "girl".
• Within France: pain au chocolat / chocolatine, serpillière / wassingue / panosse.

The varieties are described in 08. Dankova (2024, dankova-2024) compares French and Québec newspaper texts on politics and culture, about 36,000 words each. In this exploratory sample, the texts from France have more adjectives, a wider variety of them, and more participles and gerunds. The author calls the result preliminary, pending larger corpora.

6.3 Non-standard writing

Messages and social media are full of spelling variants, initialisms (mdr, tkt, jsp, slt, bjr), slang and expressive lengthening. In their review of lexical normalisation, Nishimwe, Sagot & Bawden (2024, nishimwe-2024) note that models trained on standard text lose accuracy on such content. Replacing non-standard words by standard forms is one remedy. Current models are already quite robust, though, so normalisation stays useful mainly with few resources and for sociolinguistic study.

7. False friends with English

French and English share many words that differ in sense (faux amis).

Partial false friends share one sense and differ in others: conférence (lecture or conference), expérience (experience or experiment), contrôler (to check, less often to control), pièce (room, coin, play, part), définitivement (permanently, not "definitely"), passer un examen (to sit an exam; to pass is réussir), entrée (entrance, or a starter), agenda (a diary), délai (a time limit), sympathique (likeable, not sympathetic).

Most pairs come from one Latin source and diverged after one language took the word. Sometimes French borrowed from English and changed the sense (footing, smoking; §3.3). A program that links French and English words by form alone will link wrongly.

8. How Mova handles it

What Mova has for French. There is no French lexicon in Mova. The tagger learns word forms from the training trees and guesses the rest from their letters and neighbours. On UD French GSD (UD 2.18, train → dev, 2026-10-22) it tags 96.47% of tokens correctly; the 7.2% of dev tokens it has not seen get 84.8%. French is parsed by the English parser after en xpos-fill maps the treebank's UPOS and FEATS to Penn-style tags (en/src/main.rs, xpos_of). LAS is 87.00 with Mova's own tags and 89.78 with gold tags. The English side has a full-form lexicon (en/data/forms.tsv), a lemma list (en/data/lemmas.tsv), an acronym table (en/data/acronyms.tsv) and register lexicons (global/data/register-children.tsv, register-legal.tsv, register-tale.tsv). None of these has a French counterpart, and the tagger's dictionary hint is English, so French words do not reach Mova's frames or world layers by lemma.

What a French lexicon needs, in order of what it buys:

1. Full forms with lemma and features, for unknown words and for lemmas in the frames. The candidates and their licences are listed in Sources. The Lefff and Morphalou lexicons are LGPL-LR according to their distributors, still to be confirmed. GLÀFF derives from Wiktionnaire, CC BY-SA, so it is for learning and tests only. The papers describing Lefff and GLÀFF are CC BY-NC, so we know them by their cards. The 2026 Lexique 4 article is not open access and is known by its card only.
2. Multiword expressions: the PARSEME French annotation (MWE layer CC BY 4.0 per Candito et al. 2017; check the current release before building it in) and the fixed and ExtPos annotation of UD. PolylexFLE's data licence is not stated in the paper and has to be checked.
3. Register and variety labels, from the French Wiktionary (raw dump on disk). It is CC BY-SA, so for learning and tests only, never carried into a release. The same holds for resources built on it, such as the supersense annotation of Angleraud, Barque & Candito (2025, angleraud-2025): about 85% accuracy on supersenses and 92% on the coarser hypersenses.
4. A derivational layer for the learned adjectives and the action nouns, from Démonette2 or VerNom (18 §10). The papers are CC BY 4.0, but the data licences are to be checked.
5. A French–English concept map, learned from parallel text (Europarl, the Canadian Hansards, the parallel Bibles), as tools/translate learns its interlingua. The doublets, loans and false friends of §§2, 3 and 7 show why form alone cannot give it.

What the treebanks give without a lexicon: the density of function-word expressions (fixed: 8.1 per 1,000 words in GSD, 17.3 in ParisStories) and the share of foreign material (0.49% of GSD words). Lexical MWEs (2.2% of the tokens of the PARSEME French training data, Pasquer 2017) are not in UD.

Sources for this chapter

• Ferdinand Brunot, Précis de grammaire historique de la langue française, 1899: §§141–151 (the foreign elements, the adaptation of loans), §165 (the three layers, doublets). Public domain. brunot-1899.
• Conseil supérieur de la langue française, Les rectifications de l'orthographe (1990): rule II.7 and analysis I.6 (singular and plural of loans), list F (welded foreign compounds). Official text, freely reusable. cslf-1990.
• L. Zurbuchen & R. Voigt, A Computational Analysis and Exploration of Linguistic Borrowings in French Rap Lyrics, ACL SRW 2024. CC BY 4.0. zurbuchen-2024.
• M. Candito et al., Annotation d'expressions polylexicales verbales en français, TALN 2017. CC BY 4.0. candito-2017.
• C. Pasquer, Expressions polylexicales verbales : étude de la variabilité en corpus, RÉCITAL 2017. CC BY 4.0. pasquer-2017.
• J. Bezançon et al., Forbidden FRUIT is the Sweetest: An Annotated Tweets Corpus for French Unfrozen Idioms Identification, LAW 2025. CC BY 4.0. bezancon-2025.
• S. Bagdasarov, D. Alves & E. Teich, Cheese it up: CamemBERT Outperforms Large Language Models for Identification of French Multi-word Expressions, MWE 2026. CC BY 4.0. bagdasarov-2026.
• C. Noel, N. Asher, O. Gouvert, F. Benamara & J. Hunter, EIFFEL: a novel benchmark to measure bias of English heavy training on French idiomatic expressions, ACL 2026. CC BY 4.0. noel-2026.
• Beauchemin et al., Idiom Understanding as a Tool to Measure the Dialect Gap, ACL 2026. CC BY 4.0. beauchemin-2026.
• A. Todirascu, M. Cargill & T. François, PolylexFLE : une base de données d'expressions polylexicales pour le FLE, TALN 2019. CC BY 4.0. todirascu-2019.
• F. Namer et al., Demonette2, TALN 2019. CC BY 4.0. namer-2019. D. Sanacore, N. Hathout & F. Namer, Représentation sémantique des familles dérivationnelles au moyen de frames morphosémantiques, TALN 2020. CC BY 4.0. sanacore-2020.
• N. Angleraud, L. Barque & M. Candito, Annotating the French Wiktionary with supersenses, COLING 2025. CC BY 4.0. angleraud-2025.
• N. Dankova, Lexical richness of French and Quebec journalistic texts, CLIB 2024. CC BY 4.0. dankova-2024.
• L. Nishimwe, B. Sagot & R. Bawden, Étude sur la normalisation lexicale de contenus produits par les utilisateurs, TAL 64(2), 2024. CC BY 4.0. nishimwe-2024.
• UD 2.18 French treebanks GSD, Sequoia, ParisStories, Rhapsodie, FQB, PUD, ALTS: counts of 2026-11-06 over all files of each. Quotations from ParisStories (CC BY-SA 4.0), short, with sent_id.
• Reported from general knowledge, not from a source we hold (examples our own): the loi Toubon and the 1996 terminology decree, the Québec office, dictionary register labels, the false-friend pairs, the regional words.
• Mova: en/src/main.rs (xpos_fill, xpos_of), en/data/, global/data/register-*.tsv, tools/translate; the baseline in What every language gets and the worklog of 2026-10-22; the English manual, chapter 19.

This page in Markdown → https://md.mova.elamur.ai/free/en/docs/languages/fr/19-lexicon-mwe-register/